A Longitudinal Resource for Genetic Research in Behavioral and Health Sciences Imputation Report 1000 Genomes Project reference panel (Phase 3) November 2nd, 2016 Contents I. Summary and recommendations for dbGaP users II. Study data a. Samples b. Variants c. Indel processing d. Data formatting e. Pre-phasing III. Reference panel IV. Strand alignment V. Imputation software and computing resources VI. Imputation output a. Phased output b. Genotype probabilities c. Quality metrics d. Masked variant analysis e. Downstream analysis VII. Summary VIII. References IX. Web resources X. Tables XI. Figures XII. Supplementary files 1 I. Summary and recommendations for dbGaP users Genotype imputation is the process of inferring unobserved genotypes in a study sample based on the haplotypes observed in a more densely genotyped reference sample1,2. The University of Washington Genetic Analysis Center (GAC) used IMPUTE2 software3,4 to perform genotype imputation in the Wisconsin Longitudinal Study (WLS). This report provides a detailed account of data preparation and imputation; describes the imputation output, including file formats and quality metrics; and makes recommendations for downstream analyses. Imputed results are provided as the probability of each of the three genotype states at each variant for every study participant. We recommend incorporating these imputed probabilities into any downstream analyses (e.g., as allelic dosages), rather than collapsing the probability information into the most likely genotype. Quality metrics are provided that can be used for filtering imputation results on a per-variant basis. For additional background information and a detailed description of genotype quality control (QC) on this project, please see the genotype QC report available through dbGaP (database of Genotypes and Phenotypes). II. Study data a. Samples The WLS is a long-term study of a random sample of men and women who graduated from Wisconsin high schools in 1957 and their siblings. The addition of genetic data to WLS creates opportunities for genetic studies of aging, behavior, cognition, personality, mental health, health, disease, and mortality. A total of 9,606 study samples, including duplicates, were put into genotyping production at the Center for Inherited Disease Research at Johns Hopkins University. After the GAC’s standardized QC procedures5, including resolution of sample quality and identity, genotypes are available on dbGaP for 9,012 unique WLS study participants. (Note for 15 pairs of monozygotic twins, only one member of each pair is retained in the unique set of 9,012 participants.) The GAC QC procedures yielded a set of recommended SNP and sample filters, as described in the genotype QC report. These recommended filters were then used to select study samples and SNPs for imputation. Typically we exclude from imputation samples with missing call rate (MCR) > 2%; however, no samples failed this criterion. Therefore, the imputation included all 9,012 unique genotyped study participants posted to dbGaP. For samples with gross chromosome anomalies affecting an entire chromosome, those samples were removed from the affected chromosome’s imputation but imputed for all non-affected (i.e., non- anomalous) chromosomes. As a result, one sample was excluded from chromosome 15 imputation, three from chromosome X imputation, and fifteen from imputation of the pseudo-autosomal regions PAR1 and PAR2. Note partial chromosome anomalies were addressed in post-imputation processing, described in section VI. Figure 1 shows a principal component analysis (PCA) of 9,018 unique study participants with population reference samples from the International HapMap Project6. (Note the difference between 9,018 samples in the PCA and 9,012 in the imputation is due to six participants who withdrew consent.) This PCA was done to establish ancestry orientation of study samples and presented here to visualize macro patterns of genetic ancestry. In Figure 1, study samples are color-coded by self-identified race 2 group, a variable that is intentionally not included in the dbGaP posting per WLS Institutional Review Board (IRB) stipulation. WLS participants are recruited from a relatively homogeneous population of predominately Northern and Western European ancestry. Therefore, as expected, the majority of study samples self- identify as “white” and cluster in the PCA space with HapMap reference populations of European descent. There are, however, WLS participants self-identifying as American Indian/Alaska Native, Asian, Black or African American, or as more than one race. For the most part, these samples cluster in the PCA space with the reference population(s) nearest to their self-identified race group. The IMPUTE2 algorithm discussed below recommends the use of a worldwide or “cosmopolitan” reference panel, irrespective of the genetic ancestry composition of study samples. Thus, we imputed all study samples together in one group, to the same worldwide reference panel. The local subject level identifier (“SUBJECT_ID” in annotation files) was used as the individual identifier throughout, which can be mapped to the local sample-level identifier (“SAMPLE_ID” in annotation files) using the sample-subject mapping files provided in the Supplementary Files (section XII). b. Variants This project was genotyped on the Illumina OmniExpress array, humanomniexpress-24-v1-1 annotation version A, designed to human genome build 37/hg19. For the purposes of imputation, study variants were selected using GAC-recommended quality filters described in the genotype QC report. In addition to requiring variants to pass the quality filter (“quality.filter”=TRUE), we further restricted to variants with (1) known chromosome and position (i.e., exclude unmapped variants with chromosome or position = 0); (2) located on chromosomes 1-22, X, or XY (pseudo-autosomal); and (3) with unique positions, which involved removing where “redundant”=TRUE and/or “dup.pos.disc”=TRUE. A summary of initial input variants is shown in Table 1; a list of these variants is available in the Supplementary Files. Observed genotypes, which have a maximum probability of 1, are included in the imputation output. Where an observed study variant had sporadic missing data, the missing genotypes were imputed by the pre-phasing software. Additionally, variants genotyped in the study but not used as imputation input (i.e., not passing the pre-imputation quality filters) may also appear in imputed results when available in the reference panel. This data formatting pipeline could result in discrepancies between observed genotypes posted in the primary genotype release and these imputed data. The variant annotation files accompanying this report can be used to differentiate between observed study variants used in the imputation input and the imputed variants. We refer to the former set of variants as the “imputation basis” and to the latter as the “imputation target.” These terms are analogous to the IMPUTE2 definitions of “type 2” and “type 0” variants, respectively. (Note that “type 1” variants occur only when more than one reference panel is used with IMPUTE2.) Lastly, we refer to study variants that do not occur in the reference as “study only,” or “type 3” in IMPUTE2; these are also retained in imputation output. See Figure 2 for a visual representation of these variant types. c. Indel processing The OmniExpress array contains one insertion/deletion variant, or “indel.” Indels require extra processing to prepare the study dataset for imputation to a 1000 Genomes reference panel. Specifically, 3 alleles and — in some instances — base pair positions given in the Illumina manifest must be transformed for the imputation software to identify overlap between variants in the study data and variants in the reference panel. For example, Illumina describes indel alleles with a “-“ for the deletion, e.g. “A/-,” whereas the VCF (variant call format) convention used in 1000 Genomes lists the reference base directly upstream of the deletion, e.g. “GA/G.” Harmonizing study dataset indel annotation with 1000 Genomes requires knowing the map position of indels. However, the one OmniExpress indel was unmapped in the Illumina manifest and thus not able to be harmonized. Unmapped variants are also excluded from the imputation basis, so the lack of aligned indel alleles and positions had no effect on the imputation. d. Data formatting The study genotype data were initially accessed from a binary PLINK7 file with genotypes expressed in TOP alleles. PLINK data formatting proceeded in two steps. First, at the level of the genome-wide binary PLINK file, we used genomic strand information to identify and flip the strand of SNPs where the TOP alleles were not aligned to the plus (“+”) strand of the human genome reference assembly (see section IV). In this first PLINK step we also updated monomorphic variants, where the PLINK file initially shows a “0” as one of the two alleles, to instead replace that “0” with the minor (i.e., unobserved) allele. This was done to allow the imputation software to find overlapping variants between the study and reference based on matching positions and alleles. As a second PLINK formatting step, we divided the PLINK dataset into chromosome-specific binary files. In this step, we also (1) set haploid genotypes (male chromosome X) called as heterozygotes to missing; (2) extracted study samples, i.e., removed genotyping controls and any samples with a whole chromosome anomaly on the given chromosome;